Multi-Company Panel on Building Production-Grade AI Agent Systems

abridge_/replit/_hebbia · Tech · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Production AI agents across healthcare, coding, and finance

Abridge, Replit, and Hebbia each needed high-stakes production agents that can operate reliably at scale while balancing autonomy, compliance, cost, and human trust.

Constraints:

Healthcare documentation requires clinical compliance and trust.
Coding agents need reliable long-running execution and sandbox isolation.
Financial-services agents face spiky workloads and high-value questions.
All three companies need reliability and cost optimization under production stakes.

Architecture

Across the panel, production agent systems use durable execution, tool-enabled workflows, domain-specific pipelines, and evolving autonomy controls. Abridge summarizes clinical conversations, Replit builds autonomous coding agents, and Hebbia runs document-heavy financial analysis at large agent counts.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  abridge-clinical-documentation-pipeline["Abridge Clinical Documentation Pipeline"]
  replit-coding-agent["Replit Coding Agent"]
  hebbia-financial-analysis-agents["Hebbia Financial Analysis Agents"]
  temporal-durable-execution["Temporal Durable Execution"]
  sandbox-runtime["Sandbox Runtime"]
  document-processing-pipeline["Document Processing Pipeline"]
  human-reviewers["Human Reviewers"]
  abridge-clinical-documentation-pipeline --> temporal-durable-execution
  abridge-clinical-documentation-pipeline --> human-reviewers
  replit-coding-agent --> sandbox-runtime
  replit-coding-agent --> temporal-durable-execution
  hebbia-financial-analysis-agents --> document-processing-pipeline
  hebbia-financial-analysis-agents --> temporal-durable-execution
  temporal-durable-execution --> abridge-clinical-documentation-pipeline
  temporal-durable-execution --> replit-coding-agent
  temporal-durable-execution --> hebbia-financial-analysis-agents
  sandbox-runtime --> replit-coding-agent
  document-processing-pipeline --> hebbia-financial-analysis-agents

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Abridge Clinical Documentation Pipeline`	Records, transcribes, and summarizes doctor-patient conversations.	`Temporal Durable Execution`, `Human Reviewers`
`Replit Coding Agent`	Plans, designs, writes, tests, and debugs software.	`Sandbox Runtime`, `Temporal Durable Execution`
`Hebbia Financial Analysis Agents`	Process high-value financial-services questions over large document workloads.	`Document Processing Pipeline`, `Temporal Durable Execution`
`Temporal Durable Execution`	Provides durable execution for long-running agent workflows.	`Abridge Clinical Documentation Pipeline`, `Replit Coding Agent`, `Hebbia Financial Analysis Agents`
`Sandbox Runtime`	Executes coding-agent work while isolated from the agent harness lifecycle.	`Replit Coding Agent`
`Document Processing Pipeline`	Splits and processes whole financial documents through LLMs.	`Hebbia Financial Analysis Agents`

Data flow

Domain inputs enter specialized agent pipelines: clinical conversations become notes, coding goals become planned and tested software changes, and financial questions fan out over documents. Temporal-style durable execution keeps long-running work resumable.

“Abridge tackles clinical documentation by recording and summarizing doctor-patient conversations for over 250 healthcare systems”summary, stated

“All three companies leverage Temporal for durable execution, have moved beyond proof-of-concept to production systems with high stakes”summary, stated

Why this architecture

The shared architecture reflects production maturity: durable workflows for reliability, domain-specific pipelines for high-stakes tasks, and increasing agent autonomy only where tooling and evaluation can support it.

“A common theme emerged around the evolution from constraining agents to empowering them.”summary, stated

“All three companies leverage Temporal for durable execution”summary, stated

Models used: Claude models · LLM agents · Temporal workflows

Scale

Abridge is deployed across more than 250 healthcare systems.
“over 250 major enterprise health systems including Kaiser Permanente, Sutter Health, and Mayo Clinic”summary, stated
Hebbia handles hundreds of thousands of agents for high-value financial questions.
“hundreds of thousands of agents processing high-value questions worth hundreds of millions of dollars.”summary, stated
Replit agents have increasingly long-running capabilities.
“Replit builds autonomous coding agents that can plan, design, write, test, and debug software with increasingly long-running capabilities.”summary, stated

Alternatives considered (roads not taken)

Keep agent harness inside sandbox lifecycle

Rejected because: Replit found that binding the harness to sandbox reliability caused agents to chase infrastructure failures.

“For Replit, a major breakthrough came from decoupling the agent harness from the sandbox lifecycle.”summary, stated

Use RAG to select a small context subset for financial analysis

Rejected because: Hebbia moved away from RAG because it did not deliver the answer quality customers required.

“Hebbia was early to RAG, putting it into production in 2020, but quickly moved away from it because it wasn't delivering the quality of answers their customers needed.”summary, stated

Patterns

`durable-execution` (operations)

Temporal supports long-running production workflows across all three companies.

“All three companies leverage Temporal for durable execution”summary, stated

`tool-use-orchestration` (architecture)

Agents use domain tools for clinical documentation, coding, testing, debugging, and document analysis.

“Replit builds autonomous coding agents that can plan, design, write, test, and debug software”summary, stated

`context-engineering` (context)

Hebbia's document strategy processes entire documents through LLMs rather than relying on narrow retrieved snippets.

“breaks down entire documents into chunks and runs everything through an LLM.”summary, stated

Tradeoffs

Decouple harness and sandbox

Chose Separate agent harness lifecycle from sandbox lifecycle over Running the harness inside the sandbox.

Rationale: The separation improves infrastructure reliability and prevents agents from chasing unrelated sandbox problems.

“Decoupling these components not only improved infrastructure reliability but also led to better agent outcomes”summary, stated

Process whole documents for finance

Chose Chunk and run entire documents through LLMs over RAG-selected context only.

Rationale: Customer answer quality required more complete source coverage.

“wasn't delivering the quality of answers their customers needed.”summary, stated

Gotchas (what bit them)

Infrastructure failures can mislead coding agents.

Root cause: When the agent harness was tied to sandbox lifecycle, infrastructure unreliability looked like task failure.

Resolution: Replit decoupled the agent harness from sandbox lifecycle and introduced Temporal.

“unreliable infrastructure would cause agents to try fixing problems that weren't actually related to the task at hand, wasting time and sometimes preventing successful outcomes.”summary, stated

RAG was insufficient for Hebbia's financial-answer quality bar.

Root cause: Selecting only potentially relevant context could omit material needed for high-stakes questions.

Resolution: Hebbia chunks whole documents and runs everything through an LLM as models became cheaper and more capable.

“instead of using RAG to select potentially relevant context, breaks down entire documents into chunks and runs everything through an LLM.”summary, stated

Evaluation

The panel discusses reliability, cost, model selection, clinical trust, and high-value customer outcomes but does not provide one shared benchmark.

Metrics: health systems deployed · agent count · question value · reliability

“share common challenges around reliability, cost optimization, model selection, and the evolving balance between agent autonomy and human control.”summary, stated

Outcomes

Abridge deploys clinical documentation AI across more than 250 health systems.
Replit runs increasingly long-running autonomous coding agents.
Hebbia operates financial-services agent workloads for high-value questions.

The panel spans over 250 health systems, hundreds of thousands of Hebbia agents, and high-value financial questions worth hundreds of millions of dollars.

“each building sophisticated multi-step AI pipelines that operate at significant scale with high-stakes requirements.”summary, stated